Of course by “we”, we mean educated, well-off Westerners
Artificial intelligence won’t lead to the demise of the human race but may in fact help us deal with the massive scaling up of communication that the internet has made possible and – gasp – actually help humans understand one another a little better.
Speaking at the Intel Capital conference in Palm Desert last week, three AI experts took an altogether more pragmatic and positive view of where we are going.
“An artificial intelligence machine won’t want to become a person,” argued Reza Zadeh, the CEO of Matroid. “It will be content to serve customers.”
Jean-Francois Gagne, CEO of Element AI, agreed: “We will get to systems that are self-aware but I don’t think we’ll be replicating humans – it will be different.” And a third CEO, Ben Vigoda of Gamalon, was also on board: “We can use AI as a way to bring people together.”
The fear that machines will inevitably turn against their creators says more about us as humans than it does about AI, the panelists noted. We have millions of years of evolution that has largely been about survival – the code contained in our brains is focused on making sure we get to see another day and on pass on our genes to the next generation.
But there’s no good reason why AI machines would feel the same core drive for survival. And, despite the sci-fi extrapolations like current HBO series Westworld, the truth is that AI applications are being designed to be useful at given tasks, not to exist as an autonomous being.
Peace, love and understanding
And if you want to ride that optimistic wave a little further, Vigoda of Gamalon believes that artificial intelligence could make the human race a little more understanding of itself.
Vigoda is working on a pretty intriguing new technology that is designed to take input from hundreds of thousands, even millions, of sources/people and turn it into something useful and actionable.
You could view it as the Super Alexa: a digital assistant that actually understands what you are trying to say, in the way you are trying to say it, and pulls it into an interface that can be used by a business to figure out what their customers really want.
That interface can learn over time, creating “modular AI” that starts equating to what we would see as expertise: an understanding of how things work in a specific field based on lots of inputs and queries.
In the same way that an article or a book can carry a specific idea or useful, applicable information, a modular AI could do the same, Vigoda notes.
Currently, the artificial intelligence/deep learning approaches we increasingly see in mainstream applications, like Siri, Alexa and Google Assistant are reliant on a human-created list of options that are checked or not checked in order to arrive at an understanding of what is being said.
The issue with this approach, Vigoda explained to us, is that it is not very flexible. They can’t handle uncertainty. Or work with probability. The result is most easily seen in digital assistants failing to understand context, or being unable to retain context, or personal preference.
So you can ask the system to do one thing – say, order a pizza – and unless it gets everything it needs in that order, it’s not going to be able to process it. A smarter system will recognize that some element is missing – like pick-up or delivery, or choice of pizza company – and ask for that information while also retaining the previous information.
Patterns
Critically however, a shift away from check-box deep learning and a move toward so-called “differential learning” would enable that same system to draw connections between different requests – so, for example, booking a hotel could be identified as being very similar to booking airline flights.
In this case, an AI system would start seeing things the way humans often do: a series of patterns that can be applied to other tasks, with a few tweaks here and there.
This kind of task is something we value highly in people: consultants are paid big bucks to look at a business and apply lessons from elsewhere to make it more effective or efficient. Likewise, the very best customer service reps are those that grasp what the customer wants fast, and are able to identify a persistent problem, even if it shows itself in a multitude of different, seemingly disconnected ways.
What makes Gamalon’s system particularly interesting is that it has produced a user interface that lets you see into how the system understands and makes sense of the information: it produces thought trees that can be edited.
A lot of AI systems right now often act as black boxes: you put something in and marvel at the output. Sometimes that output is not what you’re looking for – like when Microsoft’s chatbot turned into a feminist-hating Nazi in little less than a day.
Vigoda foresees his system being used by companies that field hundreds of thousands of requests from customers. The accumulating information would actually make a system smarter rather than overwhelming it, and show up patterns that an expert in the business would recognize – if they were capable of reading hundreds of thousands of messages.
He also sees staff being given variable editing rights based on their expertise: a pro who understands the business and the system could override the learning trees; a newbie could make changes but the system would quickly override them if they didn’t match with what was coming in.
